Payne College Football Power Ratings
Current Ratings · Explanation · All Predictions · Email me
The Payne College Football Power Ratings use a composite of two different systems, a Predict or margin of victory based system, and a W/L or Win/Loss only system. The ratings in each system are computed independently, and then a weighted average is used to combine the two ratings. The weighting adjusts throughout the season from using a higher weight for the Predict at the beginning of the season to a lower weight for the Predict by the end of the season. Since there is relatively little data at the beginning of the season, considering the margins of victory is needed to have sufficient data to make the ratings meaningful. By the end of the season, more data is available and teams are more rewarded for their on-the-field record than on margins of victory. Thus the overall ratings are the best system when determining how teams should be ranked based on actual game performance. However, the predict model is better when it comes to predicting the outcome of future games. The systems are defined in more detail below.
The W/L or Win/Loss ratings consider only the winner and loser of each game played. No consideration is given to margin of victory, home field advantage, or date the games are played. This system is completely BCS compliant and correlates well with the other computer ranking systems. To optimize the W/L ratings the computer determines the probability that the victor of each game played matches the actual victor based on the current ratings. It will then iterate the ratings until a maximum total probability is reached for all games played to date. The probability function is based on the P-Win concept described in more detail below. Note that for the first 4 weeks of the season (or whenever all teams become "connected"), other factors are considered to create a unique solution for the W/L model.
The Predict rankings use a predictive least squares fit of previous games played. The model used considers several factors including margin of victory, where the game is played, and when the game was played. It includes all games involving at least one FBS school. This is the model that is used for predicting all future games. To predict a game based on the Predict Ratings, the home team is predicted to win by its power rating plus the home field advantage minus the away team’s power rating (if the result is negative, then the away team is predicted to win by this margin). Click on any team name from the main ratings page, and you can see detailed predictions for that team for the regular season.
The P-Win column in the rankings and prediction page is the probability that the team will win its next game. This assumes a normal distribution of result errors and a standard deviation of 15.72. This normal distribution matches well the college football data from the last several years. The “Visitor Win Prob” on the “All Predictions” page uses a similar assumption to compute probability. These probabilities are used for the final record prediction and the Prob Undef calculations defined below.
Two models are used to predict a team’s final regular season record on the individual team pages. The probability model uses the sum of the team’s chance of winning each game as its total wins for the regular season (and computes predicted losses similarly). This is the model that is used on the main ratings page. The boolean model, just looks at each matchup and gives the team a win for each game that it is expected to win. The probability model is a better predictor of final record. The Prob Undef column is the computed probability that the team will finish the regular season undefeated.